As companies deploy AI agents in real-world operations, how to manage large numbers of agents is emerging as an issue. Amazon Web Services (AWS) and Google are drawing attention after taking different approaches.

A recent report by VentureBeat said Google focuses on managing agents at the system layer. AWS, by contrast, uses a harness approach in which developers predefine what an agent will do, which model it will use and which tools it will call, and the system runs the agent within that framework.

Speed is central to the AWS approach. When developers define an agent’s role, model and tools, Bedrock AgentCore handles the rest. Agents can run immediately with configuration settings, without building them from scratch. VentureBeat said the approach is optimised for experimentation and rapid deployment.

AWS earlier updated the Amazon Bedrock AgentCore platform and added a 'managed agent harness' function. Previously, deploying AI agents to an actual operating environment took days for developers to set up infrastructure such as computing resources, authentication protocols, persistent storage and a code execution sandbox. The managed agent harness replaces that process with writing a simple configuration file, allowing completion within minutes.

Google, through Gemini Enterprise, is focusing on providing a centralised control environment to manage multiple agents in one place. It supports central oversight of who can use which agents, which rules must be followed and what agents are doing at the moment. It is the same way Kubernetes manages hundreds of servers in one place.

When there were only a few agents inside a company and they provided only simple functions, management was not a major problem. But the situation changes when agents can run on their own for days, sending emails, modifying data and processing payments.

Long-running agents can, over time, make decisions based on outdated information or fall into confusion after receiving conflicting results from one another, a so-called state drift phenomenon. When that happens, trust in AI is bound to fall. That is why proper control is becoming as important as building quickly.